From an informatics perspective, decision-making failures in accident prevention are due to insufficient necessary safety evidence. Analyzing accident data can help in obtaining safety evidence. Currently, such a practice mostly relies on experts’ judgement and experience, which are subjective and inefficient. Furthermore, due to the inadequate safety-related theoretical support, the sustainable safety of a system can hardly be achieved purposefully. To automatically explore and obtain latent safety evidence in coal-mine data, and improve the reliability and sustainability of coal-mine safety management, a novel framework of combining data mining technology and evidence-based safety (EBS) theory is proposed, and was applied to a coal gas explosion accident. First, the term frequency-inverse document (TF-IDF) and TextRank algorithms were fused to extract keywords, and keyword evolution word cloud maps from the time dimension were drawn to obtain keyword safety evidence. Then, on the basis of the latent Dirichlet allocation (LDA) model, the best safety evidence, such as accident causation topics and causation factors, were mined, and safety decisions were given. The results show that accident data mining, based on evidence-based safety, can effectively and purposefully mine the best safety evidence, and guide safety decision making to optimize safety management models and achieve sustainable safety.